Investigating the saliency of sentiment expressions in aspect-based sentiment analysis

Anonymous

Abstract: We examine the behaviour of an aspect-based sentiment classifier built by fine-tuning the English BERT base model on the SemEval 2016 English dataset. In a set of masking experiments, we examine the extent to which the tokens which express the sentiment towards the aspect are being used by the classifier. The enhanced performance of a classifier that only sees the relevant sentiment expressions suggests that they are not being used to their full potential. Furthermore, sentiment expressions which are not directly relevant to the aspect in focus also appear to be used. We then use a gradient-based method to identify the most salient words. A comparison of these salient words, or rationales, with the sentiment expressions reveals only a moderate level of agreement. Some disagreements are related to the fixed length of the rationales and the tendency of the rationales to contain content words related to the aspect itself.

Paper Type: short